Research on coal-gangue-kaolinite identification method based on visible light and dual-energy X-ray multimodal image fusion

To address the issues of missed detections, false alarms, and limited accuracy in single-modal ore sorting under complex conditions, this paper proposes a novel classification method based on the fusion of visible light and dual-energy X-ray images. First, a laboratory setup was established to collect visible light and dual-energy X-ray images of coal, gangue, and kaolinite. The acquired high- and low-energy X-ray images were fused using an adaptive weighted fusion algorithm. Subsequently, a dual-branch backbone network was constructed to process both modalities simultaneously. Considering that visible light imaging captures surface features but fails to reveal internal composition, while X-ray imaging detects internal components but involves higher deployment complexity, we selected MobileNetV3 as the backbone. To enhance feature extraction, an Efficient Channel Attention (ECA) mechanism and a HardSwish activation function were incorporated, facilitating precise information extraction from both modalities. The extracted features were fused via a concatenation operation, leveraging the complementary advantages of the two imaging techniques to achieve real-time ore sorting. Experimental results demonstrate that the proposed fusion method significantly outperforms single-modal approaches, improving accuracy by 2.7% (over visible light) and 30.6% (over X-ray), respectively. The final recognition accuracy reaches 100%, with zero missed or false detections.